Rapid supersonic tail flame flow field calculation method based on physical information neural network

Abstract

The invention belongs to the technical field of computational fluid mechanics and artificial intelligence crossing, and particularly relates to a supersonic speed tail flame flow field rapid calculation method based on a physical information neural network, and the method comprises the following steps: S1, constructing a profile spectrum parameter data set and a tail flame flow field sample data set of a rocket engine; constructing a physical information neural network model; constructing a total loss function of a physical information neural network model, taking the profile spectrum parameter data set as an input data set of the neural network model, and normalizing the tail flame flow field data to serve as an output data set of the neural network model; training and testing the neural network model to obtain a trained neural network model; obtaining profile spectrum parameters of a rocket engine to be measured, inputting the profile spectrum parameters into the trained neural network model, and outputting a corresponding tail flame flow field distribution prediction result; the problems that an existing rocket engine tail flame flow field calculation method is high in calculation cost, physical information is lost, and large-scale efficient calculation requirements are difficult to support are solved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computational fluid dynamics and artificial intelligence, specifically involving a method for rapid calculation of supersonic exhaust flow field based on physical information neural networks. Background Technology

[0002] The exhaust plume flow field of a rocket engine is a multiphysics coupled system involving complex chemical reactions, high speed, compressibility, and turbulence. Accurate prediction of its flow field distribution is crucial for engine performance evaluation, thermal protection design, and infrared signature analysis.

[0003] Currently, high-precision computational fluid dynamics simulation is the primary means of obtaining the flow field of the exhaust plume. However, this method typically requires constructing complex turbulence and chemical reaction models and performing extremely fine mesh generation, resulting in high computational costs and long processing times, making it difficult to meet the needs of rapid iteration, real-time feedback, or extensive case screening in engineering design. In recent years, purely data-driven neural network models have been attempted for flow field prediction. While these models are fast, they are essentially "black box" models, heavily reliant on large amounts of complete training data, and exhibit poor generalization ability. For new conditions or geometric parameters not covered by the training data, their prediction results may severely violate physical laws, leading to prediction failure and insufficient reliability.

[0004] Therefore, it is necessary to invent a fast calculation method for the supersonic exhaust flow field based on a physical information neural network. Summary of the Invention

[0005] To address the problems of high computational cost, lack of physical information, and inability to support large-scale and efficient computing requirements in existing rocket engine exhaust flow field calculation methods, this invention provides a fast calculation method for supersonic exhaust flow fields based on physical information neural networks.

[0006] This invention is achieved using the following technical solution:

[0007] A fast calculation method for the supersonic exhaust plume flow field based on physical information neural networks includes the following steps:

[0008] S1: Construct a rocket engine type parameter dataset, input the type parameter dataset into a high-precision computational fluid dynamics simulation model in the existing technology, obtain the tail flame flow field data of the rocket engine under different type parameters, and construct a tail flame flow field sample dataset;

[0009] S2: Construct a physical information neural network model, whose input layer nodes correspond to the type parameters of the rocket engine, and whose output layer nodes correspond to the key physical field variables of each spatial grid point in the exhaust flow field;

[0010] S3: Construct the total loss function of the physical information neural network model, and use the spectrum parameter dataset as the input dataset of the physical information neural network model, and use the normalized tail flame flow field data as the output dataset of the physical information neural network model;

[0011] S4: Divide the input dataset and the output dataset into a training set and a test set, and use the training set and the test set to train and test the physical information neural network model to obtain a trained physical information neural network model.

[0012] S5: Obtain the type parameters of the rocket engine to be tested, input them into the trained physical information neural network model, and output the corresponding predicted results of the exhaust flow field distribution.

[0013] Furthermore, in step S1, the spectral parameters include combustion chamber temperature, combustion chamber pressure, nozzle expansion ratio, oxidant content in the propellant, and binder content.

[0014] Furthermore, in step S1, the tail flame flow field data includes pressure field, temperature field, velocity field, density field, Mach number field, and concentration fields of various components.

[0015] Furthermore, the components of the concentration field are H, O, OH, H2O, CO, CO2, HCl, H2, O2 and N2.

[0016] Furthermore, the physical information neural network model is a multilayer perceptron architecture, specifically including a feature input layer, a first fully connected layer, a first batch normalization layer, a first activation function layer, a first dropout layer, a second fully connected layer, a second batch normalization layer, a second activation function layer, a second dropout layer, a third fully connected layer, a third batch normalization layer, a third activation function layer, a third dropout layer, a fourth fully connected layer, a fourth batch normalization layer, a fourth activation function layer, and a fifth fully connected layer, connected in sequence.

[0017] The feature input layer is used to receive input feature vectors with a dimension of 5;

[0018] The fifth fully connected layer is the output layer, and its output dimension corresponds to the product of the number of spatial grid points in the exhaust flow field and the number of output physical field variables.

[0019] Furthermore, the total loss function Loss term from data fitting and physical equation residual loss term The weighted composition is expressed as follows:

[0020] ;

[0021] in, and These are the balancing weighting coefficients.

[0022] Furthermore, the data fitting loss term The mean square error is the difference between the predicted tensor output by the physical information neural network model and the corresponding true value output by the high-precision computational fluid dynamics simulation model.

[0023] Furthermore, the residual loss term of the physical equation Calculate using the following steps:

[0024] (a) Some physical field variables are extracted from the prediction tensor output by the physical information neural network model. The physical field variables include pressure field p, velocity field u in the x direction, velocity field v in the y direction, and density field ρ.

[0025] (b) Based on the preset spatial coordinates and grid spacing corresponding to the spatial grid points, the spatial gradient of the partial physical field variables is calculated using automatic differentiation technology;

[0026] (c) Based on the partial physical field variables resolved in step (a) and the spatial gradient calculated in step (b), calculate the residual of the hydrodynamic conservation control equations used to describe the tail flame flow field. The hydrodynamic conservation control equations include the continuity equation and the momentum equation.

[0027] (d) The sum of the squares of the residuals of the continuity equation and the momentum equation is averaged over all spatial grid points to obtain the residual loss term of the physical equation. .

[0028] Furthermore, in step S4, during the training of the physical information neural network model, a weight adjustment strategy is used to adjust the balancing weight coefficients. and ;

[0029] The weight adjustment strategy is as follows: based on the data fitting loss term... and the residual loss term of the physical equation During training, adjust the weight coefficients of each loss as the magnitude or gradient changes to keep the contribution of each loss to the total loss function on the same order of magnitude.

[0030] Compared with existing methods for calculating exhaust plume flow fields, this invention provides a fast calculation method for supersonic exhaust plume flow fields based on physical information neural networks, which has the following advantages:

[0031] First, high computational efficiency: Once the physical information neural network model is trained, the prediction process only requires one forward propagation calculation, which takes milliseconds to seconds. Compared with computational fluid dynamics simulation, which takes several days, the speed is improved by several orders of magnitude, realizing second-level flow field prediction and greatly accelerating the design iteration process.

[0032] Secondly, physical consistency is guaranteed: by directly embedding the residuals of the fluid dynamics conservation control equations (such as the continuity equation and momentum equation) as physical constraints into the loss function, the model's prediction results strictly follow the basic physical conservation laws. Even in regions where the training data is sparse or missing, it can give physically reasonable predictions, overcoming the physical fallacies that may occur in purely data-driven models.

[0033] Third, strong generalization ability and data efficiency: The introduction of physical equations provides the model with a powerful inductive bias, enabling it to better extrapolate to spectrum parameter data not fully covered by the training data. At the same time, combined with the weight adjustment strategy, the training process is optimized, further reducing the dependence on a large amount of complete computational fluid dynamics simulation data.

[0034] Fourth, efficient end-to-end mapping: The model establishes a direct and fast mapping from engine design spectrum parameters to the full flow field distribution, eliminating the complex meshing and iterative solution process in traditional simulation, simplifying the prediction process, and making it easy to integrate into existing design optimization platforms to achieve efficient analysis and decision-making.

[0035] In summary, this invention effectively solves the problems of high computational cost, lack of physical information, and difficulty in supporting large-scale and efficient computation in existing exhaust flow field calculation methods, and is applicable to supersonic exhaust flow fields. Attached Figure Description

[0036] Figure 1 This is the overall flowchart of the present invention.

[0037] Figure 2 This is a graph of the loss function during the training process in an embodiment of the present invention.

[0038] Figure 3 This is a comparison chart of the prediction results of the method of the present invention and the calculation results of high-precision computational fluid dynamics in the embodiments. Detailed Implementation

[0039] The present invention will be further explained and described below with reference to the accompanying drawings, embodiments, and comparative examples. Obviously, the described embodiments are only a part of the embodiments, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Example

[0041] A fast calculation method for the supersonic exhaust flow field based on physical information neural networks, such as Figure 1 As shown, it includes the following steps:

[0042] S1: Construct a dataset of rocket engine type parameters. These type parameters include the combustion chamber temperature T. C Combustion chamber pressure P C Nozzle expansion ratio ε, and oxidizer AP content Y in the propellant. AP and the content of adhesive HTPB Y HTPB .

[0043] The aforementioned model parameter dataset is input into a high-precision computational fluid dynamics (CFD) simulation model in the prior art to perform numerical simulation, thereby obtaining the exhaust flow field data of the rocket engine under different model parameters, and thus constructing an exhaust flow field sample dataset.

[0044] The exhaust flow field data includes key physical field variables at all spatial grid points, specifically: pressure field P, temperature field T, velocity field U, density field R, Mach number field Ma, and concentration fields Y of 10 components. s And the concentration field Y s The 10 components are H, O, OH, H2O, CO, CO2, HCl, H2, O2, and N2. This step establishes a mapping relationship from spectral parameters to full flow field data, i.e., (T C P C , ε, Y AP Y HTPB ) → (P, T, U, R, Ma, Y s ).

[0045] S2: Construct a physical information neural network model. In this embodiment, the model adopts a multilayer perceptron architecture. Its input layer is designed with 5 nodes, corresponding to the 5 spectral parameters (T) in step S1. C P C , ε, Y AP Y HTPB The output layer nodes correspond to the key physical field variables of each spatial grid point in the exhaust flow field.

[0046] The specific architecture of this multilayer perceptron neural network model includes a feature input layer, a first fully connected layer, a first batch normalization layer, a first activation function layer, a first dropout layer, a second fully connected layer, a second batch normalization layer, a second activation function layer, a second dropout layer, a third fully connected layer, a third batch normalization layer, a third activation function layer, a third dropout layer, a fourth fully connected layer, a fourth batch normalization layer, a fourth activation function layer, and a fifth fully connected layer, all connected in sequence.

[0047] The feature input layer is used to receive input feature vectors with a dimension of 5.

[0048] The first activation function layer, the second activation function layer, the third activation function layer, and the fourth activation function layer all use the ReLU activation function.

[0049] The fifth fully connected layer is the output layer, and its output dimension (i.e., the total number of neurons) corresponds to the product of the total number of spatial grid points in the exhaust flow field and the total number of output physical field variables. In this embodiment, the total number of output physical field variables is 16, specifically including pressure, temperature, velocity components, density, Mach number, and the concentrations of 10 components (H, O, OH, H2O, CO, CO2, HCl, H2, O2, and N2). The spatial grid is 50×100, and the output dimension of this layer is 50×100×16, mapping high-dimensional features to specific spatial distributions and physical quantities.

[0050] S3: The total loss function for constructing the physical information neural network model This is used to guide the training of the model. The total loss function consists of the data fitting loss term. and physical equation residual loss term The weighted composition is expressed as follows:

[0051] ;

[0052] in, and These are the balancing weighting coefficients.

[0053] Meanwhile, the spectrum parameter dataset constructed in step S1 is used as the input dataset of the physical information neural network model, and the corresponding tail flame flow field data (i.e., CFD simulation results) is normalized and used as the output dataset (i.e., training labels) of the neural network.

[0054] The data fitting loss term The mean square error is the difference between the predicted tensor output by the physical information neural network model and the corresponding true value output by the high-precision CFD simulation model.

[0055] The physical equation residual loss term The following steps are used to calculate and embed the fluid dynamics conservation laws as soft constraints into the network:

[0056] (a) Analyze the flow field variables: From the prediction tensor output by the physical information neural network model, analyze some of the physical field variables required to calculate the physical residual, including the pressure field p, the velocity field u in the x direction, the velocity field v in the y direction, and the density field ρ.

[0057] (b) Calculate the spatial gradient: Based on the preset spatial coordinates (x, y) and grid spacing Δx, Δy corresponding to the points in the flow field grid, calculate the spatial gradient of each physical field variable in step (a), such as the pressure gradient, using automatic differentiation techniques. (Quantity) and velocity gradient , .

[0058] (c) Calculate the residuals of the equations: Based on the partial physical field variables analyzed in step (a) and the spatial gradient calculated in step (b), calculate the residuals of the hydrodynamic conservation control equations used to describe the tail flame flow field. The hydrodynamic conservation control equations include the continuity equation and the momentum equation.

[0059] Continuity equation residual R cont Represented as .

[0060] Momentum equation residual R mom Represented as .

[0061] (d) Convergence physical loss: The sum of the squares of the residuals of the continuity equation and the momentum equation is averaged over all spatial grid points to obtain the residual loss term of the physical equation. :

[0062] ;

[0063] Where N is the total number of spatial grid points.

[0064] S4: Divide the input dataset and the output dataset prepared in step S3 into a training set and a test set. Use the training set to train the physical information neural network model and use the test set to evaluate its performance to obtain the trained physical information neural network model.

[0065] During training, a weight adjustment strategy is used to adjust the balancing weight coefficients. and This strategy is based on and During training, adjust the weights of the two losses according to changes in numerical magnitude or gradient, so that the two losses are proportional to the total loss function. The contributions of each component remain on the same order of magnitude, thus resolving the training imbalance problem in multi-objective optimization and promoting model convergence. The training process minimizes the total loss function using the gradient descent algorithm. To optimize network parameters.

[0066] S5: For new, unpredictable rocket engine operating conditions, obtain their type spectrum parameters (T). C PC , ε, Y AP Y HTPB The data is then input into the physical information neural network model trained in step S4. The model, through a single forward propagation calculation, instantly (in milliseconds to seconds) outputs the predicted tail flame flow field distribution covering all spatial grid points, including detailed data on all key physical field variables.

[0067] Figure 2 This is a schematic diagram of the training loss function of the physical information neural network in this invention. Figure 2 It can be seen that the model's total loss function converges steadily during the training process.

[0068] Figure 3 This is a schematic diagram comparing the calculation results of key physical quantities of the exhaust flow field under different spectral parameters of the present invention and existing high-precision CFD methods. In the diagram, PINN-PLUME represents the method based on physical information neural networks described in this invention. Figure 3 It can be seen that the present invention exhibits better robustness and adaptability on the exhaust flow field datasets under different conditions, and can accurately calculate the exhaust flow field of rocket engines under different morphological parameters.

[0069] Experiments show that traditional CFD simulations take approximately 10 hours per run, while the method of this invention, after training, only requires about 0.1 seconds for a single prediction, resulting in a computational efficiency improvement of approximately 3.6 × 10⁻⁶. 5 This significantly improves the computational efficiency of rocket engine exhaust flow field calculations.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fast calculation method for the supersonic exhaust plume flow field based on a physical information neural network, characterized in that: Includes the following steps: S1: Construct a rocket engine type parameter dataset, input the type parameter dataset into a high-precision computational fluid dynamics simulation model in the existing technology, obtain the tail flame flow field data of the rocket engine under different type parameters, and construct a tail flame flow field sample dataset; S2: Construct a physical information neural network model, whose input layer nodes correspond to the type parameters of the rocket engine, and whose output layer nodes correspond to the key physical field variables of each spatial grid point in the exhaust flow field; S3: Construct the total loss function of the physical information neural network model, and use the spectrum parameter dataset as the input dataset of the physical information neural network model, and use the normalized tail flame flow field data as the output dataset of the physical information neural network model; S4: Divide the input dataset and the output dataset into a training set and a test set, and use the training set and the test set to train and test the physical information neural network model to obtain a trained physical information neural network model. S5: Obtain the type parameters of the rocket engine to be tested, input them into the trained physical information neural network model, and output the corresponding predicted results of the exhaust flow field distribution.

2. The method for rapid calculation of supersonic exhaust plume flow field based on physical information neural network according to claim 1, characterized in that: In step S1, the spectral parameters include combustion chamber temperature, combustion chamber pressure, nozzle expansion ratio, oxidant content in the propellant, and binder content.

3. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 1, characterized in that: In step S1, the exhaust flow field data includes pressure field, temperature field, velocity field, density field, Mach number field, and concentration field of various components.

4. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 3, characterized in that: The concentration field consists of H, O, OH, H2O, CO, CO2, HCl, H2, O2, and N2.

5. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 1, characterized in that: The physical information neural network model is a multilayer perceptron architecture, specifically including a feature input layer, a first fully connected layer, a first batch normalization layer, a first activation function layer, a first dropout layer, a second fully connected layer, a second batch normalization layer, a second activation function layer, a second dropout layer, a third fully connected layer, a third batch normalization layer, a third activation function layer, a third dropout layer, a fourth fully connected layer, a fourth batch normalization layer, a fourth activation function layer, and a fifth fully connected layer, connected in sequence. The feature input layer is used to receive input feature vectors with a dimension of 5; The fifth fully connected layer is the output layer, and its output dimension corresponds to the product of the number of spatial grid points in the exhaust flow field and the number of output physical field variables.

6. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 1, characterized in that: The total loss function Loss term from data fitting and physical equation residual loss term The weighted composition is expressed as follows: ； in, and These are the balancing weighting coefficients.

7. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 6, characterized in that: The data fitting loss term The mean square error is the difference between the predicted tensor output by the physical information neural network model and the corresponding true value output by the high-precision computational fluid dynamics simulation model.

8. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 6, characterized in that: The physical equation residual loss term Calculate using the following steps: (a) Some physical field variables are extracted from the prediction tensor output by the physical information neural network model. The physical field variables include pressure field p, velocity field u in the x direction, velocity field v in the y direction, and density field ρ. (b) Based on the preset spatial coordinates and grid spacing corresponding to the spatial grid points, the spatial gradient of the partial physical field variables is calculated using automatic differentiation technology; (c) Based on the partial physical field variables resolved in step (a) and the spatial gradient calculated in step (b), calculate the residual of the hydrodynamic conservation control equations used to describe the tail flame flow field. The hydrodynamic conservation control equations include the continuity equation and the momentum equation. (d) The sum of the squares of the residuals of the continuity equation and the momentum equation is averaged over all spatial grid points to obtain the residual loss term of the physical equation. .

9. The method for rapid calculation of supersonic exhaust flow field based on physical information neural network according to claim 6, characterized in that: In step S4, during the training of the physical information neural network model, a weight adjustment strategy is used to adjust the balancing weight coefficients. and ; The weight adjustment strategy is as follows: based on the data fitting loss term... and the residual loss term of the physical equation During training, adjust the weight coefficients of each loss as the magnitude or gradient changes, so that the contribution of each loss to the total loss function remains on the same order of magnitude.

Images